Symbiont AI and Embodied Symbiotic Learning

Abstract

Various approaches have been explored to enable AI agents to assist human users in complex tasks. To date, these systems have been narrowly designed to tackle specific rather than general tasks, are brittle when taken outside of controlled environments, or can only be trained by technically proficient users. In this paper we propose a new type of general-purpose assistive agent called a Symbiont AI that is designed to support users with greater task flexibility and environmental resilience, and that can be trained on new tasks by non-technical users using a novel method we call Embodied Symbiotic Learning. Instead of programming an AI to perform tasks (either by explicit coding, demonstration, or machine learning) and then deploying the AI to assist a human, we form a Human-AI Symbiotic System in which an AI partners with a human and learns to assist the human in real-time, while the human learns to share the workload with the AI. Each partner develops a theory of mind of the other in real-time, so a partnership-specific set of expectations and communication norms emerges through the shared interactions. This allows the AI to be useful to the human even in open-ended or underspecified tasks. We also consider practical design considerations for such systems and present experimentally testable predictions that follow from this theory.